PPO-clip limits policy updates by clipping the policy probability ratio inside a surrogate objective. It addresses reinforcement-learning instability when repeatedly updating from experiences collected by an older policy. section 12.1.5
The method belongs to proximal policy optimization, alongside PPO-penalty. Whereas PPO-penalty discourages large changes with a KL-divergence penalty inspired by TRPO, PPO-clip limits the update directly through the objective. section 12.1.5
Clipped surrogate objective
A plotted clipped surrogate objective makes the update limit visible: policy changes beyond the permitted ratio range stop improving the clipped contribution.
